On 28/10/2020 18:22, Thomas Gleixner wrote:
On Wed, Oct 28 2020 at 20:33, John Garry wrote:
Hi Thomas,
+int irq_update_affinity_desc(unsigned int irq,+ struct irq_affinity_desc *affinity) +{ + unsigned long flags; + struct irq_desc *desc = irq_get_desc_lock(irq, &flags, 0); + + if (!desc) + return -EINVAL;Just looking at it some more. This needs a check whether the interrupt is actually shut down. Otherwise the update will corrupt state. Something like this: if (irqd_is_started(&desc->irq_data)) return -EBUSY; But all of this can't work on x86 due to the way how vector allocation works. Let me think about that.
Is the problem that we reserve per-cpu managed interrupt space when allocated irq vectors on x86, and so later changing managed vs non-managed setting for irqs messes up this accounting somehow?
Cheers, John
